Pith. sign in

REVIEW 2 major objections 8 minor 3 cited by

Light3R-SfM: Towards Feed-forward Structure-from-Motion

T0 review · 2 major / 8 minor · reviewed 2026-08-10 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A learned attention module replaces global optimization in SfM, cutting runtime by 49x.

desk verdict Genuine step toward feed-forward SfM with strong empirical results, but the complexity analysis in Section 3.1 is arithmetically wrong and the paper overstates its theoretical scalability case. read the letter →

arxiv 2501.14914 v1 pith:BGKA2NK6 submitted 2025-01-24 cs.CV cs.LG

Sven Elflein , Qunjie Zhou , Sérgio Agostinho , Laura Leal-Taixé This is my paper · ORCID
classification cs.CVcs.LG
keywords feed-forwardstructure-from-motionlatentglobalalignmentattentionmechanismscenegraphshortestpathtreepointmapregressioncameraposeestimation3Dreconstruction
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Light3R-SfM claims that the iterative global optimization at the heart of classical and learning-based structure-from-motion can be replaced by a learnable attention mechanism that aligns images in feature space. If the claim holds, reconstructing an unordered photo collection becomes a single feed-forward pass: encode all images, let latent global tokens share context across every view, decode only N-1 pairs along a retrieval-built shortest-path tree, and merge the pairwise pointmaps with closed-form Procrustes alignments. The payoff is a >49x speedup over the optimization-based MASt3R-SfM pipeline (33 seconds versus about 27 minutes for a 200-image scene) while keeping pose accuracy competitive on Tanks&Temples, CO3Dv2, and Waymo. This matters because runtime, not accuracy, is often the practical bottleneck for SfM in applications such as driving scenes, large photo collections, and downstream neural rendering.

What carries the argument

The central object is the latent global alignment module: for each image, average-pool its dense tokens into one global token, then alternate self-attention over the N global tokens with cross-attention from each image's dense tokens to the global set, in L=4 blocks, with a residual connection from the original tokens. This factorizes attention cost from O((N*T)^2) to O($N^{2}$ + N*T), where T is tokens per image, making global context sharing feasible for large collections. The second mechanism is the shortest path tree (SPT) scene graph: pairwise cosine similarities of average-pooled encoder embeddings define edge costs, and the SPT connects all images with exactly N-1 edges while keeping paths shallow to limit drift during accumulation. The third is global accumulation: traversing the SPT breadth-first, each newly reached node is registered into the global frame by a confidence-weighted Procrustes alignment computed in closed form, so no iterative solver is ever run.

What would settle it

Take a scene whose two halves overlap through exactly one image pair, and make that pair's average-pooled token similarity artificially low (e.g., heavy motion blur on the bridge image). If the reconstruction splits into two internally consistent but mutually misaligned pieces, the retrieval assumption fails in a single controlled experiment. A quantitative version: correlate pairwise cosine similarity with ground-truth overlap across a diverse dataset; the method's reliability rests on that correlation being strong specifically on the edges the SPT selects.

Watch

Extended reading notes

Core claim

The paper establishes that a factored attention module—self-attention among average-pooled global tokens for all images, followed by cross-attention from dense per-image tokens to those updated global tokens—can globally align camera poses and pointmaps implicitly in latent space, so that pairwise decoders emit pointmaps in a common coordinate frame without any iterative alignment or bundle adjustment. The authors combine this with a scene graph built as a retrieval-score-guided shortest path tree, needing only N-1 pairwise decodings, and merge local pointmaps by traversing the tree in breadth-first order with confidence-weighted Procrustes alignments. They report that the method rivals optimization-based pipelines (within a few accuracy points of MASt3R-SfM at tight 5-degree thresholds on Tanks&Temples, and roughly matching it at about 195x lower runtime on the Waymo driving benchmark) while clearly surpassing the memory-bank feed-forward baseline Spann3R in both accuracy and speed. The authors also state explicitly that their model does not yet scale to tens of thousands of images and that tight-threshold pose accuracy still trails optimization-based methods, likely due to the low resolution of learned feature tokens.

Load-bearing premise

The scene graph is built only from cosine similarities of average-pooled image tokens, so if those similarities fail to connect truly overlapping views (or connect non-overlapping ones), the tree edges are wrong and the global reconstruction breaks into pieces that are internally consistent but globally misaligned; the paper's own failure cases show exactly this.

Editorial extensions

If this is right

  • Reconstructing a 200-image collection takes about 33 seconds instead of about 27 minutes, so SfM becomes practical for near-real-time and runtime-constrained applications; the paper reports up to about 198x speedups on full sequences.
  • The pipeline accepts unordered image sets directly; unlike the memory-bank baseline Spann3R, it does not need temporal ordering or an exhaustive pairwise ordering search, which lets it scale to sequences of roughly 1100 images without out-of-memory errors.
  • Pose accuracy at loose thresholds (10-30 degrees) is close to optimization-based methods, suggesting the coarse poses are already usable for downstream tasks such as Gaussian splatting; the authors note that a small optimization stage on top could close the tight-threshold gap.
  • Ablations show the latent alignment module adds about 7 points of RRA@5 over the pairwise-only baseline, and the SPT graph outperforms an MST by about 15 points of RRA@5, indicating both components are load-bearing.
  • The learned confidence maps let users trade registration completeness for pose accuracy at inference by raising a confidence threshold, which the paper demonstrates across thresholds 3 to 7.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Because retrieval similarity alone decides the graph, I would expect the method to be fragile on scenes with repeated or ambiguous structure (tunnels, corridors, facades), where pooled-token cosine similarity is a weaker overlap signal; a testable extension is to add geometric verification or uncertainty-aware graph construction.
  • The authors' own oracle-graph ablation shows that replacing the learned retrieval with ground-truth overlap scores yields only a small gain, which suggests the accuracy ceiling is currently set by the pairwise decoder and resolution rather than by graph construction.
  • Since accumulation is a deterministic tree traversal, worst-case drift is bounded by the depth of the SPT; re-rooting the tree or averaging several SPT hypotheses could harden the method without changing the architecture.
  • The tight-threshold gap the authors attribute to low resolution implies a concrete test: train or evaluate at higher token resolution; if the gap persists at matched resolution, the limitation lies in the latent alignment itself rather than in image resolution.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

2 major / 8 minor

Summary. Light3R-SfM proposes a feed-forward SfM pipeline that replaces global optimization with a learned latent alignment module. Given an unordered image collection, the method encodes each image, exchanges information between images via self-attention over per-image global tokens followed by cross-attention from image tokens to those global tokens, builds a sparse scene graph via a shortest-path tree from average-pooled features, decodes pairwise pointmaps for each edge, and accumulates them via closed-form Procrustes alignment. The paper reports competitive pose accuracy on Tanks&Temples, CO3Dv2, and Waymo, with a claimed >49x speedup over MASt3R-SfM on 200-view scenes and superior accuracy and speed relative to Spann3R. The paper also provides ablations for the latent alignment, global supervision, scene graph construction, and hyperparameters.

Significance. If the empirical results hold, Light3R-SfM is a meaningful advance toward optimization-free SfM: it demonstrates that a learned attention module in latent space can replace iterative global alignment for multi-view pose estimation, at a fraction of the runtime. The experimental evaluation is extensive (three datasets, multiple view settings, comparisons to both optimization-based and feed-forward baselines), and the ablations are careful, including an Oracle scene-graph study that quantifies the gap to a perfect retrieval. The paper is transparent about its limitations, notably the accuracy gap at tight thresholds and the failure mode of retrieval-driven graph construction. However, the theoretical complexity analysis in Sec. 3.1 is incorrect, and the Procrustes convention in Sec. 3.3 is ambiguous; these need correction before publication.

major comments (2)
  1. [Sec. 3.1, Eqs. (2)-(3) and Discussion] The claimed time complexity of the latent global alignment module, stated as O(N^2 + N*T), is incorrect. In each latent alignment block, the cross-attention in Eq. (3) lets each of the T tokens of an image attend to all N global tokens, costing O(N*T) per image and O(N^2*T) across N images. Adding the global self-attention O(N^2) gives O(N^2*T) per block, not O(N^2+N*T). Furthermore, the comparison with naive self-attention O((N*T)^2)=O(N^2*T^2) is not "the same asymptotic complexity class"; the two expressions differ by a factor of T. If T grows with image resolution, the method scales quadratically in N times T, not near-linearly as the formula and the surrounding text suggest. The empirical speedups at N <= 1100 are not invalidated, but the theoretical basis for scaling to much larger collections is overstated and should be corrected or rephrased.
  2. [Sec. 3.3, Eqs. (6)-(7)] The direction of the Procrustes transformation P_k is undefined. If Procrustes(X_k, X_{k,k}) returns the transformation that aligns the second point set to the first (i.e., from the local frame of node k to the global frame), then Eq. (7) should apply P_k to X_{k,l}, not P_k^{-1}. If instead P_k maps from global to local, then the inverse is appropriate, but this convention must be stated explicitly. As written, the reader cannot determine whether the accumulation correctly maps node l into the global coordinate frame, which is central to the reconstruction procedure.
minor comments (8)
  1. [Eq. (4)] The notation S_ij = ⟨∥¯F_i∥2, ∥¯F_j∥2⟩ is mathematically meaningless because the norms are scalars, and the scalar product of two scalars is not defined; presumably the intended expression is the cosine similarity between the normalized embeddings, ⟨¯F_i/∥¯F_i∥, ¯F_j/∥¯F_j∥⟩.
  2. [Sec. 3.3, Eq. (6)] The range specification "log C_k ∈ [0, ∞]^{H×W}" appears inconsistent with a confidence map whose values lie in [0,1], since the logarithm of such values is non-positive. Please clarify the scaling or the intended range.
  3. [Sec. 4.4, Table 5 and Fig. 11] The Oracle scene-graph study shows that the retrieval-based SPT is only 0.4 RRA@5 points below Oracle (Table 5), which supports the graph construction, but the paper provides no quantitative measure of retrieval failures (e.g., the proportion of missing edges or disconnected components) on the evaluated datasets; the failure mode shown in Fig. 11 is only qualitative.
  4. [Table 1, full row for COLMAP] The entry "GT" in the COLMAP row for the full-sequence setting is unexplained; please clarify whether this indicates that COLMAP is used to generate the ground-truth poses and is therefore not evaluated, or something else.
  5. [Sec. 4.4, Table 6] The sentence "leading to decreased rejected frames yet overall more accurate poses" is contradicted by the table, which shows the registration rate decreasing as the confidence threshold increases, meaning more frames are rejected; the wording should be corrected to "increased rejected frames" or "lower registration rate".
  6. [Sec. 4.1] The statement that the method is "competitive with other learning-based methods including VGGSfM" is not fully supported by Table 1 at 100 and 200 views, where VGGSfM leads by 7.4 and 32.1 points in RRA@5, respectively; if VGGSfM's per-dataset training (mentioned in the CO3Dv2 section) explains the gap, this should be disclosed for the Tanks&Temples comparison.
  7. [General] The paper does not release code or trained weights, which limits independent verification of the reported runtimes and accuracies, especially the headline >49x speedup claim; the authors should consider releasing the code or providing a detailed reproducible protocol.
  8. [Typos] There are minor typographical issues, such as "(iiv)" in the pipeline description in Sec. 3 and "learning rale" in Appendix A; these should be corrected.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the method is validated on held-out benchmarks and its components are defined independently of the reported outputs.

full rationale

Light3R-SfM's derivation chain is self-contained and empirically grounded. The latent alignment module, scene graph, and global accumulation are defined by explicit equations (Eqs. 1-7) that do not take the predicted poses or reconstructions as inputs. Global tokens are mean-pooled encoder tokens; the SPT is built from cosine similarities of average-pooled image tokens; pairwise pointmaps are decoded from the aligned tokens; and final poses are extracted from pointmaps via fixed PnP/Procrustes operations. The training losses (Eqs. 8-11) supervise pointmaps against ground-truth poses and pointmaps from the training data, with a Procrustes alignment that only removes gauge freedom; no test-set labels are used to set constants. The ablations and confidence-threshold selection are model selection on evaluation subsets, which is standard practice rather than circular reasoning. No load-bearing self-citation chain exists: citations to prior work (DUSt3R, MASt3R, CoTracker, Spann3R) are baselines, pretrained initializations, or inspiration, and none is invoked as an unverified uniqueness theorem. The paper candidly states limitations (does not scale to tens of thousands of images, poses lag at tight thresholds, retrieval failures cause disjoint sub-reconstructions), which are consistent with the evaluation and do not indicate that any prediction reduces to its inputs. The Sec. 3.1 complexity claim (O(N^2 + N*T) versus O((N*T)^2) and calling them the same asymptotic class) is arithmetically questionable, since cross-attention over all N global tokens for each image yields O(N^2*T) per block; this is a correctness/analysis issue in the scalability argument, not a circularity, because the runtime formula is not used as an input to the method's derivation.

Assumptions & free parameters 4 free parameters · 6 assumptions · 0 invented entities

The contribution is empirical and architectural. No new physical entities are introduced. The main free parameters are hyperparameters chosen by ablation. The method relies on pretrained MASt3R/DUSt3R weights as a strong geometric prior and on standard closed-form geometric solvers (Procrustes, PnP) for assembling poses.

free parameters (4)
  • Weight of global supervision lambda = 0.1
    Chosen via ablation on Tanks and Temples 100-view subset (supp Table 9); higher values (1.0) reduced accuracy.
  • Number of latent global alignment layers L = 4
    Chosen via ablation in supp Table 8; L=8 showed diminishing returns and higher memory cost.
  • Confidence threshold at inference = 3 (or 90% quantile)
    Set by hand in supplementary Appendix A; Table 6 shows accuracy improves with higher thresholds at the cost of registration rate.
  • Training graph size N = 8
    Selected due to GPU memory; supp Table 11 shows N=10 gives slightly better results but exceeds memory for the high-resolution model.
assumptions (6)
  • standard math Closed-form least-squares Procrustes alignment (Umeyama) provides the optimal rigid transformation between two matched point clouds.
    Used in Sec 3.3 Eq (6) to register each new node to the global reconstruction during tree traversal.
  • standard math RANSAC-PnP with a robust focal estimator extracts a camera pose from a predicted pointmap and confidence map.
    Used at inference to convert globally aligned pointmaps into camera poses (supp Appendix A).
  • domain assumption A single average-pooled global token per image is a sufficient summary for sharing multi-view context across all images.
    Central to the latent global alignment module in Sec 3.1; no proof is given that this low-dimensional bottleneck preserves enough information for accurate global alignment.
  • domain assumption Cosine similarity between average-pooled image tokens reliably identifies image pairs with sufficient visual overlap for scene graph construction.
    Used in Sec 3.2 to build the shortest path tree; the paper's own failure cases (Fig 11) show that retrieval failures cause globally inconsistent sub-reconstructions.
  • domain assumption Tree-structured accumulation with N-1 edges and confidence-weighted Procrustes is sufficient to avoid fatal drift across large image collections.
    The method in Sec 3.3 assumes that error does not accumulate excessively along the tree; the MST ablation (Table 5) shows drift sensitivity, motivating the SPT but not eliminating drift in general.
  • domain assumption Training on mostly static scenes (Waymo, CO3Dv2, MegaDepth, TartanAir) transfers to in-the-wild scenes including some dynamic content.
    Supp Appendix C notes that dynamic objects degrade reconstruction quality because the model assigns high confidence to dynamic regions; this is acknowledged as a limitation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Light3R-SfM: Towards Feed-forward Structure-from-Motion." pith.science (2026). https://pith.science/paper/BGKA2NK6

@misc{pith2026250114914,
  author       = {Pith},
  title        = {Pith review of: Light3R-SfM: Towards Feed-forward Structure-from-Motion},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BGKA2NK6}},
  note         = {Machine review of arXiv:2501.14914}
}
read the original abstract

We present Light3R-SfM, a feed-forward, end-to-end learnable framework for efficient large-scale Structure-from-Motion (SfM) from unconstrained image collections. Unlike existing SfM solutions that rely on costly matching and global optimization to achieve accurate 3D reconstructions, Light3R-SfM addresses this limitation through a novel latent global alignment module. This module replaces traditional global optimization with a learnable attention mechanism, effectively capturing multi-view constraints across images for robust and precise camera pose estimation. Light3R-SfM constructs a sparse scene graph via retrieval-score-guided shortest path tree to dramatically reduce memory usage and computational overhead compared to the naive approach. Extensive experiments demonstrate that Light3R-SfM achieves competitive accuracy while significantly reducing runtime, making it ideal for 3D reconstruction tasks in real-world applications with a runtime constraint. This work pioneers a data-driven, feed-forward SfM approach, paving the way toward scalable, accurate, and efficient 3D reconstruction in the wild.

Figures

Figures reproduced from arXiv: 2501.14914 by the authors.

Figure 1
Figure 1. Processing speed vs. accuracy for various SfM meth￾ods. Our work significantly decreases the runtime across vari￾ous sizes of image collections compared to traditional pipelines while obtaining comparable accuracy. Results are measured on the Tanks&Temples dataset. an unconstrained stereo 3D reconstruction model through pointmap regression, i.e., by directly predicting 3D points in a common reference system for ever… view at source ↗
Figure 2
Figure 2. Light3R-SfM Pipeline. Given an unordered set of images, we first encode them to obtain image tokens from which we average pool global features for constructing a shortest path tree. We next feed image tokens into our attention-based latent global alignment to enable global context sharing. Afterwards, for each edge in the SPT, we decode pairwise pointmaps using the implicitly aligned feature tokens. Finally, we use … view at source ↗
Figure 3
Figure 3. Qualitative comparison on a Waymo scene. Note how the MASt3R-SfM reconstruction does not truthfully reconstruct the 90° turn, while Spann3R predictions degrade after tens of frames. second clips recorded at 10Hz from an autonomous vehicle. For each sequence, we use the 200 input frames from the forward-looking camera. Baselines. We compare against our concurrent works Spann3R and MASt3R-SfM, which also build on top … view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: Reconstructing opposite-oriented cameras. After con￾ditioning Light3R-SfM’s decoder with the output from our global latent alignment, it is able to predict pointmaps even for images recorded in opposite directions, suggesting the latent global align￾ment has learned a …
Figure 6
Figure 6. Figure 6: Global confidence map (right) produced by Light3R [PITH_FULL_IMAGE:figures/full_fig_p013_6.png]
Figure 5
Figure 5. Figure 5: CDF of pose errors on 100-view Tanks&Temples scenes. thresholds between feed-forward approaches (Light3R￾SfM, Spann3R [48]) and optimization-based approaches, however, this gap rapidly shrinks for our method when mov￾ing towards looser thresholds, while Spann3R is cons…
Figure 7
Figure 7. Figure 7: CDF of per-scene 3D reconstruction errors. RRA@5 ↑ RTA@5 ↑ ATE ↓ L 2 33.1 35.5 0.033 4 35.7 36.9 0.032 8 35.3 36.9 0.032 [PITH_FULL_IMAGE:figures/full_fig_p014_7.png]
Figure 8
Figure 8. Figure 8: Qualitative examples of reconstruction of Tanks & Temples scenes. Delivery area Kicker Playground Terrace Office [PITH_FULL_IMAGE:figures/full_fig_p016_8.png]
Figure 9
Figure 9. Figure 9: Qualitative examples of reconstruction of ETH3D scenes. 16 [PITH_FULL_IMAGE:figures/full_fig_p016_9.png]
Figure 10
Figure 10. Figure 10: More comparisons on Waymo. Comparing from left-to-right: MASt3R-SfM, Spann3R, Light3R-SfM. 17 [PITH_FULL_IMAGE:figures/full_fig_p017_10.png]
Figure 11
Figure 11. Figure 11: Failure cases on the Tanks & Temples dataset. 18 [PITH_FULL_IMAGE:figures/full_fig_p018_11.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 3 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. CuSfM: CUDA-Accelerated Structure-from-Motion

    cs.CV 2025-10 conditional novelty 5.0 of 10

    A CUDA-accelerated SfM system reports faster offline mapping than COLMAP and better trajectory accuracy when initialized with odometry-like poses, though the comparison is not fully fair.

  2. DeblurSplat: SfM-free 3D Gaussian Splatting with Event Camera for Robust Deblurring

    cs.CV 2025-09 conditional novelty 5.0 of 10

    A pose-free deblurring 3D Gaussian Splatting pipeline using DUSt3R point clouds, confidence-balanced sampling, and event-decoded latent image supervision.

  3. Reconstructing 4D Spatial Intelligence: A Survey

    cs.CV 2025-07 accept novelty 4.0 of 10

    A review that classifies 4D scene reconstruction methods into five progressive levels: low-level cues, scene components, dynamic scenes, interactions, and physics.

Reference graph

Works this paper leans on

60 extracted references · 47 canonical work pages · cited by 3 Pith papers

  1. [1]

    Building rome in a day

    Sameer Agarwal, Yasutaka Furukawa, Noah Snavely, Ian Si- mon, Brian Curless, Steven M Seitz, and Richard Szeliski. Building rome in a day. Communications of the ACM , 54 (10):105–112, 2011. 2

    2011

  2. [2]

    Global motion estimation from point matches

    Mica Arie-Nachimson, Shahar Z Kovalsky, Ira Kemelmacher-Shlizerman, Amit Singer, and Ronen Basri. Global motion estimation from point matches. In 2012 Second international conference on 3D imaging, modeling, processing, visualization & transmission , pages 81–88. IEEE, 2012. 2

    work page 2012

  3. [3]

    Mip-nerf: A multiscale representation for anti-aliasing neu- ral radiance fields

    Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neu- ral radiance fields. In Proceedings of the IEEE/CVF inter- national conference on computer vision , pages 5855–5864,

  4. [4]

    H. G. Barrow, J. M. Tenenbaum, R. C. Bolles, and H. C. Wolf. Parametric correspondence and chamfer matching: Two new techniques for image matching. In Proceedings of the 5th International Joint Conference on Artificial Intel- ligence - Volume 2, pages 659–663. 13

    work page

  5. [5]

    Dsac-differentiable ransac for camera localization

    Eric Brachmann, Alexander Krull, Sebastian Nowozin, Jamie Shotton, Frank Michel, Stefan Gumhold, and Carsten Rother. Dsac-differentiable ransac for camera localization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6684–6692, 2017. 2

    work page 2017

  6. [6]

    Accelerated coordinate encoding: Learning to re- localize in minutes using rgb and poses

    Eric Brachmann, Tommaso Cavallari, and Victor Adrian Prisacariu. Accelerated coordinate encoding: Learning to re- localize in minutes using rgb and poses. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5044–5053, 2023

    work page 2023

  7. [7]

    Scene coordinate reconstruction: Pos- ing of image collections via incremental learning of a relo- calizer

    Eric Brachmann, Jamie Wynn, Shuai Chen, Tommaso Cav- allari, ´Aron Monszpart, Daniyar Turmukhambetov, and Vic- tor Adrian Prisacariu. Scene coordinate reconstruction: Pos- ing of image collections via incremental learning of a relo- calizer. In European Conference on Computer Vision, 2024. 1, 2, 5, 19

    work page 2024

  8. [8]

    A pose-only solution to visual reconstruction and navi- gation

    Qi Cai, Lilian Zhang, Yuanxin Wu, Wenxian Yu, and Dewen Hu. A pose-only solution to visual reconstruction and navi- gation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(1):73–86, 2021. 1, 2

    work page 2021

Show all 60 references
  1. [9]

    Global structure-from-motion by similarity averaging

    Zhaopeng Cui and Ping Tan. Global structure-from-motion by similarity averaging. In Proceedings of the IEEE Inter- national Conference on Computer Vision , pages 864–872,

  2. [10]

    Superpoint: Self-supervised interest point detection and description

    Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabi- novich. Superpoint: Self-supervised interest point detection and description. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops , pages 224–236, 2018. 2, 5

    2018

  3. [11]

    An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Syl- vain Gelly, Jakob Uszkoreit, and Neil Houlsby. An Image is Worth 16x16 Words: Transformers for Image Recognition ...

    2020

  4. [12]

    Mast3r- sfm: a fully-integrated solution for unconstrained structure- from-motion

    Bardienus Duisterhof, Lojze Zust, Philippe Weinzaepfel, Vincent Leroy, Yohann Cabon, and Jerome Revaud. Mast3r- sfm: a fully-integrated solution for unconstrained structure- from-motion. arXiv preprint arXiv:2409.19152, 2024. 1, 2, 3, 5, 6, 7, 19

    2024 arXiv

  5. [13]

    Dkm: Dense kernelized feature matching for geometry estimation

    Johan Edstedt, Ioannis Athanasiadis, M ˚arten Wadenb ¨ack, and Michael Felsberg. Dkm: Dense kernelized feature matching for geometry estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17765–17775, 2023. 2

    2023

  6. [14]

    Roma: Robust dense feature matching

    Johan Edstedt, Qiyu Sun, Georg B ¨okman, M ˚arten Wadenb¨ack, and Michael Felsberg. Roma: Robust dense feature matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 19790–19800, 2024. 2

    2024

  7. [15]

    Fischler and Robert C

    Martin A. Fischler and Robert C. Bolles. Random sample consensus: A paradigm for model fitting with applications to image analysis and automated cartography. Commun. ACM, (6):381–395, 1981. 12

    1981

  8. [16]

    L. R. Ford. Network Flow Theory.RAND Corporation, Santa Monica, CA, 1956. 3

    1956

  9. [17]

    Build- ing rome on a cloudless day

    Jan-Michael Frahm, Pierre Fite-Georgel, David Gallup, Tim Johnson, Rahul Raguram, Changchang Wu, Yi-Hung Jen, Enrique Dunn, Brian Clipp, Svetlana Lazebnik, et al. Build- ing rome on a cloudless day. In European Conference on Computer Vision, pages 368–381. Springer, 2010. 2

    2010

  10. [18]

    Detector-free struc- ture from motion

    Xingyi He, Jiaming Sun, Yifan Wang, Sida Peng, Qixing Huang, Hujun Bao, and Xiaowei Zhou. Detector-free struc- ture from motion. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 21594–21603, 2024. 2, 5

    2024

  11. [19]

    Co- Tracker: It is Better to Track Together

    Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co- Tracker: It is Better to Track Together. In European Confer- ence on Computer Vision (ECCV), 2024. 3

    2024

  12. [20]

    3d gaussian splatting for real-time radiance field rendering

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimk ¨uhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139–1,

  13. [21]

    Tanks and temples: Benchmarking large-scale scene reconstruction

    Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics, 36(4), 2017. 5, 6, 7, 8, 12

    2017

  14. [22]

    Ground- ing image matching in 3d with mast3r.European Conference on Computer Vision, 2024

    Vincent Leroy, Yohann Cabon, and J´erˆome Revaud. Ground- ing image matching in 3d with mast3r.European Conference on Computer Vision, 2024. 1, 2, 3, 7, 12, 15

    2024

  15. [23]

    MegaDepth: Learning Single- View Depth Prediction From Internet Photos

    Zhengqi Li and Noah Snavely. MegaDepth: Learning Single- View Depth Prediction From Internet Photos. In Proceed- ings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2041–2050, 2018. 12 9

    2018

  16. [24]

    Relpose++: Recovering 6d poses from sparse-view ob- servations

    Amy Lin, Jason Y Zhang, Deva Ramanan, and Shubham Tul- siani. Relpose++: Recovering 6d poses from sparse-view ob- servations. arXiv preprint arXiv:2305.04926, 2023. 6, 7

    2023 arXiv

  17. [25]

    Pixel-perfect structure-from- motion with featuremetric refinement

    Philipp Lindenberger, Paul-Edouard Sarlin, Viktor Lars- son, and Marc Pollefeys. Pixel-perfect structure-from- motion with featuremetric refinement. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5987–5997, 2021. 1, 2, 5, 7

    2021

  18. [26]

    Lightglue: Local feature matching at light speed

    Philipp Lindenberger, Paul-Edouard Sarlin, and Marc Polle- feys. Lightglue: Local feature matching at light speed. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17627–17638, 2023. 2

    2023

  19. [27]

    An efficient and robust hybrid sfm method for large-scale scenes

    Zhendong Liu, Wenhu Qv, Haolin Cai, Hongliang Guan, and Shuaizhe Zhang. An efficient and robust hybrid sfm method for large-scale scenes. Remote Sensing, 15(3):769, 2023. 2

    2023

  20. [28]

    Decoupled Weight Decay Regularization, 2019

    Ilya Loshchilov and Frank Hutter. Decoupled Weight Decay Regularization, 2019. 12

    2019

  21. [29]

    Nerf: Representing scenes as neural radiance fields for view syn- thesis

    Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view syn- thesis. Communications of the ACM , 65(1):99–106, 2021. 1

    2021

  22. [30]

    Global structure-from-motion revisited

    Linfei Pan, D ´aniel Bar ´ath, Marc Pollefeys, and Johannes L Sch¨onberger. Global structure-from-motion revisited. In Eu- ropean Conference on Computer Vision (ECCV), 2024. 1, 2, 3, 5, 7, 19

    2024

  23. [31]

    Rethinking depth estimation for multi- view stereo: A unified representation

    Rui Peng, Rongjie Wang, Zhenyu Wang, Yawen Lai, and Ronggang Wang. Rethinking depth estimation for multi- view stereo: A unified representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8645–8654, 2022. 1

    2022

  24. [32]

    Com- mon Objects in 3D: Large-Scale Learning and Evaluation of Real-Life 3D Category Reconstruction

    Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Com- mon Objects in 3D: Large-Scale Learning and Evaluation of Real-Life 3D Category Reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vi...

    2021

  25. [33]

    R2d2: Reliable and repeatable detec- tor and descriptor

    Jerome Revaud, Cesar De Souza, Martin Humenberger, and Philippe Weinzaepfel. R2d2: Reliable and repeatable detec- tor and descriptor. Advances in neural information process- ing systems, 32, 2019. 2

    2019

  26. [34]

    From coarse to fine: Robust hierarchical localization at large scale

    Paul-Edouard Sarlin, Cesar Cadena, Roland Siegwart, and Marcin Dymczyk. From coarse to fine: Robust hierarchical localization at large scale. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 12716–12725, 2019. 1

    2019

  27. [35]

    Superglue: Learning feature matching with graph neural networks

    Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4938–4947, 2020. 2, 5

    2020

  28. [36]

    Improving image-based localization by active correspondence search

    Torsten Sattler, Bastian Leibe, and Leif Kobbelt. Improving image-based localization by active correspondence search. In Computer Vision–ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part I 12, pages 752–765. Springer...

    2012

  29. [37]

    Structure- from-motion revisited

    Johannes L Schonberger and Jan-Michael Frahm. Structure- from-motion revisited. In Proceedings of the IEEE con- ference on computer vision and pattern recognition , pages 4104–4113, 2016. 1, 2, 3, 5, 7

    2016

  30. [38]

    Schonberger, Silvano Galliani, Torsten Sattler, Konrad Schindler, Marc Pollefeys, and An- dreas Geiger

    Thomas Schops, Johannes L. Schonberger, Silvano Galliani, Torsten Sattler, Konrad Schindler, Marc Pollefeys, and An- dreas Geiger. A Multi-View Stereo Benchmark With High- Resolution Images and Multi-Camera Videos. In Proceed- ings of the IEEE Conference on Computer Vision and...

    2017

  31. [39]

    Scene co- ordinate regression forests for camera relocalization in rgb-d images

    Jamie Shotton, Ben Glocker, Christopher Zach, Shahram Izadi, Antonio Criminisi, and Andrew Fitzgibbon. Scene co- ordinate regression forests for camera relocalization in rgb-d images. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2930–...

    2013

  32. [40]

    Flowmap: High-quality camera poses, in- trinsics, and depth via gradient descent

    Cameron Smith, David Charatan, Ayush Tewari, and Vin- cent Sitzmann. Flowmap: High-quality camera poses, in- trinsics, and depth via gradient descent. arXiv preprint arXiv:2404.15259, 2024. 1, 2, 5

    2024 arXiv

  33. [41]

    Photo tourism: exploring photo collections in 3d

    Noah Snavely, Steven M Seitz, and Richard Szeliski. Photo tourism: exploring photo collections in 3d. In ACM siggraph 2006 papers, pages 835–846. 2006. 1, 2

    2006

  34. [42]

    Loftr: Detector-free local feature matching with transformers

    Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. In Proceedings of the IEEE/CVF con- ference on computer vision and pattern recognition , pages 8922–8931, 2021. 2

    2021

  35. [43]

    Scalability in Perception for Autonomous Driving: Waymo Open Dataset

    Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, Vijay Vasudevan, Wei Han, Jiquan Ngiam, Hang Zhao, Aleksei Timofeev, Scott Et- tinger, Maxim Krivokon, Amy Gao, Aditya Joshi, Yu Zhan...

    2020

  36. [44]

    A Consistently Fast and Globally Optimal Solution to the Perspective-n-Point Problem

    George Terzakis and Manolis Lourakis. A Consistently Fast and Globally Optimal Solution to the Perspective-n-Point Problem. In Computer Vision – ECCV 2020, pages 478–494, Cham, 2020. 12

    2020

  37. [45]

    S. Umeyama. Least-squares estimation of transformation parameters between two point patterns. IEEE Transactions on Pattern Analysis and Machine Intelligence, (4):376–380,

  38. [46]

    Attention is All you Need

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszko- reit, Llion Jones, Aidan N Gomez, Ł ukasz Kaiser, and Illia Polosukhin. Attention is All you Need. In Advances in Neu- ral Information Processing Systems, 2017. 12

    2017

  39. [47]

    Wong, and Chen-Yi Lee

    Sheng Wan, Tung-Yu Wu, Wing H. Wong, and Chen-Yi Lee. Confnet: Predict with Confidence. In 2018 IEEE Interna- tional Conference on Acoustics, Speech and Signal Process- ing (ICASSP), pages 2921–2925, 2018. 5

    2018

  40. [48]

    3d reconstruction with spatial memory

    Hengyi Wang and Lourdes Agapito. 3d reconstruction with spatial memory. arXiv preprint arXiv:2408.16061, 2024. 2, 3, 7, 13, 19

    2024 arXiv

  41. [49]

    Posediffusion: Solving pose estimation via diffusion-aided 10 bundle adjustment

    Jianyuan Wang, Christian Rupprecht, and David Novotny. Posediffusion: Solving pose estimation via diffusion-aided 10 bundle adjustment. In Proceedings of the IEEE/CVF Inter- national Conference on Computer Vision, pages 9773–9783,

  42. [50]

    Vggsfm: Visual geometry grounded deep structure from motion

    Jianyuan Wang, Nikita Karaev, Christian Rupprecht, and David Novotny. Vggsfm: Visual geometry grounded deep structure from motion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 21686–21697, 2024. 1, 2, 5, 7, 12, 19

    2024

  43. [51]

    Dust3r: Geometric 3d vi- sion made easy

    Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vi- sion made easy. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20697– 20709, 2024. 1, 2, 4, 5, 7, 12, 15

    2024

  44. [52]

    TartanAir: A Dataset to Push the Limits of Visual SLAM

    Wenshan Wang, Delong Zhu, Xiangwei Wang, Yaoyu Hu, Yuheng Qiu, Chen Wang, Yafei Hu, Ashish Kapoor, and Se- bastian Scherer. TartanAir: A Dataset to Push the Limits of Visual SLAM. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4909–4916,

    2020

  45. [53]

    CroCo v2: Improved Cross-view Completion Pre- training for Stereo Matching and Optical Flow

    Philippe Weinzaepfel, Thomas Lucas, Vincent Leroy, Yohann Cabon, Vaibhav Arora, Romain Br ´egier, Gabriela Csurka, Leonid Antsfeld, Boris Chidlovskii, and Jerome Re- vaud. CroCo v2: Improved Cross-view Completion Pre- training for Stereo Matching and Optical Flow. In Proceed- ...

    2023

  46. [54]

    Weiszfeld

    E. Weiszfeld. Sur le point pour lequel la Somme des dis- tances de n points donn´es est minimum. Tohoku Mathemati- cal Journal, First Series, pages 355–386, 1937. 12

    1937

  47. [55]

    Robust global translations with 1dsfm

    Kyle Wilson and Noah Snavely. Robust global translations with 1dsfm. In Computer Vision–ECCV 2014: 13th Eu- ropean Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part III 13, pages 61–75. Springer, 2014. 1, 2

    2014

  48. [56]

    Towards linear-time incremental struc- ture from motion

    Changchang Wu. Towards linear-time incremental struc- ture from motion. In 2013 International Conference on 3D Vision-3DV 2013, pages 127–134. IEEE, 2013. 1, 2

    2013

  49. [57]

    Image retrieval for structure-from-motion via graph convolutional network

    Shen Yan, Maojun Zhang, Shiming Lai, Yu Liu, and Yang Peng. Image retrieval for structure-from-motion via graph convolutional network. Information Sciences , 573:20–36,

  50. [58]

    Monst3r: A simple approach for estimat- ing geometry in the presence of motion

    Junyi Zhang, Charles Herrmann, Junhwa Hur, Varun Jam- pani, Trevor Darrell, Forrester Cole, Deqing Sun, and Ming- Hsuan Yang. Monst3r: A simple approach for estimat- ing geometry in the presence of motion. arXiv preprint arXiv:2410.03825, 2024. 3

    arXiv

  51. [59]

    Patch2pix: Epipolar-guided pixel-level correspondences

    Qunjie Zhou, Torsten Sattler, and Laura Leal-Taixe. Patch2pix: Epipolar-guided pixel-level correspondences. In Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 4669–4678, 2021. 2 11 Supplementary Material for Light3R-SfM In this supple...

    2021

  52. [60]

    By default we use a threshold of 3, or the 90%-quantile if all confidences fall below the threshold

    and first estimate the focal length with a robust estima- tor [54] and then proceed to extract the pose with RANSAC- PnP [15, 44] from points with their corresponding confi- dence in C i larger than a threshold. By default we use a threshold of 3, or the 90%-quantile if all co...

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.